Abstract
Rapid and accurate diagnosis of COVID-19 is critical for pandemic management. While RT-PCR remains the diagnostic gold standard, medical imaging provides a faster alternative for clinical assessment. This paper presents a novel hybrid radiomic framework, GMM-CNN, which leverages Gaussian Mixture Models (GMMs) to encode feature distributions learned by deep convolutional neural networks. By applying a Random Forest classifier to these encoded features, the proposed method achieves superior performance in distinguishing COVID-19 from other pneumonias in both CT and X-ray images, reaching an Area Under the Curve (AUC) of up to 99.45%.
Introduction
The COVID-19 pandemic highlighted the urgent need for efficient diagnostic tools. Expert radiologists often find it challenging to differentiate COVID-19 from other viral or bacterial pneumonias due to overlapping visual features, such as ground-glass opacities (GGO). While standard deep learning (CNN) models show promise, they typically require massive datasets to avoid overfitting and can be difficult to interpret. Conversely, traditional radiomics may lack the flexibility to capture complex disease signatures. This study introduces a middle ground: a robust, automated system that captures tissue heterogeneity through statistical modeling of deep features.
Methodology
The proposed GMM-CNN pipeline consists of four main stages:
- Feature Extraction: Pre-trained CNN architectures, specifically ResNet50 and DarkNet-19, serve as feature extractors, providing raw feature maps from multiple layers.
- GMM Encoding: Instead of using raw maps directly, the distribution of values within these maps is modeled using a Gaussian Mixture Model (GMM). This captures the "multiscale texture" and heterogeneity of the lung tissue.
- Feature Representation: Each feature map is represented by a vector of GMM parameters, including weights, means, and variances.
- Classification: These parameters are fed into a Random Forest (RF) classifier, chosen for its ability to handle high-dimensional data and provide better interpretability compared to standard neural networks.
The core innovation lies in modeling the distribution of features. As shown in the figure above, while COVID and non-COVID images may share similar feature modes, their statistical weights and positions differ significantly, providing a unique "radiomic signature" for the disease.
Experimental Setup and Datasets
The model was evaluated on a large dataset comprising 5,254 axial CT images and 8,084 X-ray images. The dataset included both COVID-19 cases and non-COVID controls (including other respiratory infections) to ensure high specificity. The researchers found that a small number of Gaussian components ($k=2$ or $3$) was optimal; increasing the number of components beyond this led to overfitting as the model began to capture noise rather than generalizable patterns.
Results and Discussion
The GMM-CNN approach significantly outperformed standard end-to-end CNN fine-tuning across all tested image modalities. Key performance metrics include:
- Accuracy: Achieved 96.00% – 96.70% across different folds.
- AUC-ROC: Reached an impressive 99.29% – 99.45%, indicating extremely high reliability.
- Performance Boost: The hybrid method improved accuracy by 5–6% and boosted the AUC-ROC by approximately 3% compared to using standard CNNs directly.
As illustrated in the AUC-ROC curves, the hybrid GMM-CNN models (GMM-ResNet and GMM-DarkNet) consistently achieved higher sensitivity and specificity compared to their standalone counterparts. Statistical analysis (Chi-square tests) confirmed that these improvements were significant ($p < 0.05$).
Conclusion
The GMM-CNN framework provides a powerful tool for COVID-19 diagnosis, effectively capturing the heterogeneity of lung tissue through statistical encoding of deep features. This approach offers a robust alternative to standard deep learning, particularly when data is limited or high interpretability is required. Future work will focus on integrating clinical metadata (e.g., age, symptoms) and exploring the model's utility in predicting disease severity and patient survival outcomes.
Summary updated